The central argument is that WebRTC’s real-time design can undermine voice AI. Built for human conversation, it favors low latency: audio retransmissions are generally disabled, and its jitter buffer may not wait for recovered packets. A brief connection outage can therefore erase part of a user’s prompt, leaving a speech recognizer or language model with incomplete context and potentially wasting tokens or producing a guess. The author argues that AI agents can process audio faster than people speak, so they have time to recover missing packets before the model needs the complete prompt. WebRTC also complicates generated speech: it sends audio at human playback speed and largely plays it according to arrival time, rather than using timestamps to buffer and schedule it.
The proposed alternative is Media over QUIC (MoQ), which can retransmit audio for several seconds and send generated speech ahead of playback, storing it in a timestamped buffer. Because compressed audio is small, the author argues, this can improve reliability at little bandwidth or memory cost. The piece acknowledges WebRTC’s strengths, including its maturity and jitter handling, but contends that Google’s control of the browser implementation makes audio improvements slow and difficult to adopt. It calls for fixes to WebRTC’s audio timing and buffering, while advocating MoQ, now supported by Pipecat. The demonstration is limited: other services in the audio pipeline can still lose packets, and interruptions require a way to clear buffered speech.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.